I went to Rootscamp. I heard a lot of frustration and heartache and shock from folks who participated in 2016. Their stories woke up some old memories (and cautionary tales), which I’m sharing here hoping they help us plot a positive path forward.
Part 1: Maryland
“I can hear you! I can hear you! The rest of the world hears you! And the people who knocked these buildings down will hear all of us soon!”
— President George W. Bush, September 14, 2001
The 2004 campaign began for me on November 6, 2002: The day after my boss, Kathleen Kennedy Townsend, lost her bid to become the first woman governor of Maryland.
2002 was a 9/11-era sweep that put law-and-order Republicans into office at all levels of government. Many Democratic constituencies either flipped parties or simply stayed home. In Maryland, we lost by about 67,000 votes — close enough that any twenty of a hundred different decisions, made the other way, would have flipped to the outcome.
In a postmortem analysis, we saw historically-strong Democratic areas continue to vote Democrat by a strong margin, but waning support in so-called “purple” neighborhoods. Democratic support in strong Republican neighborhoods fell even further. That is, precincts that had supported Democrats in the past at 70% may have dropped a couple points to support our campaign around 68%, but historically-35% Democratic precincts dropped 10 points further, to around 20%.
This result matched our investment: we put more time and money per Democratic potential voter in areas where we had a lot of support, and spent less time and money trying to retain the votes of those Democratic supporters whose neighbors tended to vote for Republicans. (And I’m not talking about the “Democratic” voters who register in Maryland as Democrats to have a vote in primaries but always vote Republican in the general elections.)
I was in charge of targeting, so in theory I could have done something to protest or alter the approach we used. But I understood this strategy delivered the White House for Bill Clinton in 1992 and again in 1996. We tried the same thing for Al Gore and it worked in enough places to win the popular vote (but not the electoral college), so even in a bad-looking year, it would probably work to put liberal hero Bobby Kennedy’s daughter into the governor’s mansion in a state that hadn’t had a Republican governor since Spiro Agnew. I went with the strategy because I didn’t know any better and the logic of voter-inertia made sense.
The problem with assuming historic party support and turnout average %s would hold steady, is that it logically requires all aspects of a campaign to be on par with the averages of its predecessors. Scientists say “ceteris paribus” meaning “other things equal,” to describe conditions under which expectations may hold. In this case, “history will repeat itself as long as nothing changes.” We tend to think of the electorate in terms of swing voters and stable bases. But if the opponent is well-organized and his/her party has never been well-organized before, then we need to start questioning history-based turnout assumptions in opposition strongholds. Or if we’re not able to open as many field offices or recruit as many volunteers as our predecessors, then we need to tweak turnout assumptions among those voters in borderline areas we can’t afford to engage. In 2002, conditions were very different from past races, but we hadn’t done the hard work of becoming aware of the whole field and digging deep into the political engagement history before setting our expectations and plan.
Before the election, the most frequent feedback we received on phones and at the doors during GOTV was an expression of frustration that we never came around except just before an election. Even in state leg districts where full-time constituent services operations were strong, goodwill didn’t seem to travel up the ticket. For some of us, the takeaway was for top-ticket campaigns to start having real conversations with voters early. And maybe just don’t ever stop. In 2002, internet-based centralized voter contact was just starting to take shape. This new technology would soon enable constant constituent management efforts.
The weather on election day was wet and cold. I spent most of the day huddled up in the back of a crumbling Baltimore rowhouse at 25th and Charles, but I went out to the polls and watched voters abandon voting lines as it got dark and the temperature dropped into the low 40s under a light rain. We got about a half inch in Baltimore and in Montgomery County, and a little more in Prince George’s County. There’s nothing we could have done to prevent the weather, but we could have built a “rainy day buffer” into our vote goals and campaign plan. Academic research has subsequently shown rain on election day tends to lower voter turnout by 1% and boost GOP vote share by 2.5% per 1" precipitation. We didn’t know these figures in 2002, but now we know to pad our targets. I always advise adding an extra 5–10% to vote goals to account for unknowns, like the weather. As a friend who used to compete in track and field reminds me, “you gotta run through the tape.”
We knew pretty early in the day it was going to be a squeaker if we were to win, and we spent the whole afternoon watching the Weather Channel’s local updates. As the big green swirl traveled east on the live satellite map, our chances of holding the Governor’s mansion swept down the drain like so much rainwater.
The day after the election, our offices were robbed. As we were packing up, thieves ducked in and made off with the unopened stash of cigars and liquor we had procured for election-night celebrations, along with my winter coat. Though it was just a simple crime of opportunity, it felt as personal and nasty as getting hit with a bank overdraft fee on a cup of coffee.
I spent the last few dollars of my last paycheck in an effort to revive a fading campaign relationship. It didn’t work out, and the rest of my cold, coat-less autumn was full of odd jobs, interviews, calls and emails chasing work for the new year. I worked my way on to the radar of the lead consultant for Howard Dean’s presidential exploratory effort. I offered myself up as a volunteer, and spent Christmas with family in Iowa hoping Dean’s people would invite me to come aboard. Finally, I got a green-light. I loaded up my car with computers abandoned by the Maryland Democratic Party, and moved to Burlington, Vermont.
I prayed my credit card would hold out until I got a paying gig.
Part 2: Vermont and Iowa
“I’m Howard Dean and I’m here to represent the Democratic wing of the Democratic party.”
— Governor Howard Dean, February 21, 2003
My station wagon lurched into Burlington in a blizzard a week after my 22nd birthday. I had two elections and a redistricting cycle worth of political data-management under my belt, and was also pretty good at fixing computers and setting up networks. I immediately set to making sure all the office machines were running properly. A wise colleague of mine likes to say he can assess the basic infrastructural health of any office organization by asking everyone in it to print a document from the public internet — I spent my first week in Vermont creating this ability for my new co-workers.
Having done it a couple times now, I can say there’s few professional experiences as exciting or character-forging as being among the first couple dozen staff on a presidential campaign. A million things need to get done, comfort zones be damned. To be successful, everyone must tuck egos away and go do. And we must take the time to get the foundational pieces right, because in this same “get-stuff-done” effort, the person who comes after you may not check the foundation before they build new stuff on top.
We established three important technology platforms in these early days: First, we created the capacity to accept online contributions, setting up our merchant account and negotiating discount rates with banks. For some it may be hard to imagine an era when this process took weeks and not minutes. A brilliant colleague with experience in online commerce led this effort, enabling our candidate to run the first “people-powered” campaign in the new McCain-Feingold era. Second, we set up our NGP finance system, compiling and consolidating the database of high-dollar donors we would still need to support our campaign, even as we pursued our innovative low-dollar strategy. Third, we acquired the Iowa and New Hampshire State Democratic Party voter files to facilitate voter contact in the caucus and primary. In 2003, common wisdom held the winner of these contests would become the nominee.
To access the list of caucus-goers with their addresses and phone numbers and prior vote and caucus history, the Iowa party (IDP) required candidates to buy into a new shared internet-based tool called the Voter Activation Network or VAN. Before VAN, I had only received voter files in a raw data format (like a several-million-row spreadsheet) via FTP, or on physical media like a CD or Zip disk. Unlike NGP, where our data was firewalled from their other clients in a self-contained database and server, in the VAN all the caucus campaigns’ data were technically in the same database, programmatically-segregated based on user permissions.
While VAN was innovative and promising, I was skeptical and suspicious of this new technology in a primary setting. I perceived the shared platform as an espionage risk — that is, I thought the Kerry or Edwards campaigns might be able to obtain our precious voter contact information due to some unforeseen glitch. Because the platform was untested and I was paranoid, I recommended using VAN as a means to receive base data updates (like phone number updates, new records, moved/deceased flags) from the IDP, but to manage the data in a separate web-based system that we controlled ourselves.
In this effort to avoid the risk of industrial espionage, I exposed the nascent Dean campaign to a much bigger liability that would ultimately contribute to its collapse.
About a month into my Vermont tenure, the committee still hadn’t raised much money. A new leadership team started to dig in and displace some of the original staff. In the face of uncertainty, I accepted a job offer to become a tech-focused union organizer, headquartered in Washington. I packed my car again — minus the dozen computers I brought with me — and headed south. On the abnormally-warm March day I left Burlington, I remember Lake Champlain groaning, and UVM kids playing beer pong shirtless among piles of snow.
I cheered Dr. Dean’s field-leading second-quarter fundraising through a soggy D.C. spring. I followed the Sleepless Summer Tour through searing August heat while staffing a referendum in Albuquerque. Dean appeared to be far outpacing rivals for the nomination.
As I made plans to spend holidays with my family in Iowa, I reached out to my old campaign colleagues to learn how I could help out around HQ in Des Moines. The campaign’s numbers looked very positive, but they needed a little help tweaking their reporting system and getting contact data together in advance of the Perfect Storm — an organizing blitz attracting thousands of out-of-state volunteers in the weeks before the Caucus.
A few days before Christmas, I borrowed my dad’s car and headed to the campaign’s Iowa HQ. It was about a 45 minute drive down I-35 to Des Moines with the setting sun glaring at me from the southwest. Rush hour congestion in central Iowa was barely noticeable after a few months of driving in DC.
1408 Locust Street was a large and labyrinthine property originally built as a Cadillac dealership. As a nerdy Iowa high-schooler in the late 1990s, I made regular trips there — it had been home to the Midwest Computer Supply Store. Instead of stacks of games and books as in the old days, the building was now full of activists making phone calls and stacks of campaign literature. The multitude of scattered extension cords, phone lines, and computers in various states of disrepair looked like they could have been left behind by the former tenants. A staffer led me through the main showroom, up a back staircase and into the small wood-paneled office where the state leadership were holed up.
As I walked into the small upstairs bunker, a milestone was announced: the campaign could count over 130,000 Iowans as supporters. Those running the call said this meant more people had committed to caucus for Dean in January than had turned out in total to caucus in 1988, the high-water-mark for Democratic caucus participation. Staff were told Iowa was in the bag, the nomination all but clinched.
The State and Field Directors had been my colleagues on Gore 2000. They trusted me and asked their voter file manager to give me administrative access to their online database and also to their VAN. Even though we expected an overwhelming victory, leadership wanted me to make sure everything was ship-shape. I took a seat on a dusty old sofa, fired up my laptop, and got to work. I would spend about 60 hours on this sofa over the next 4 days, running a thorough internal audit on the campaign’s database.
I identified three critical data-integrity issues. Even though it had been nine months since I last worked full-time for the campaign, two of the three were my fault — direct results of my recommendation to steer clear of VAN as an operational database.
First issue was of recency: The link between the VAN and the campaign’s voter database was not automatic, and while voters’ phone numbers were being added and updated in the VAN in real-time by Edwards, Kerry, and Gephardt staffers, this data had to be manually pulled from VAN and pushed into our campaign’s system. These sync tasks were frequent, but not constant, and there was some efficiency lost as the two systems were out-of-sync. (Though arguably, the other campaigns weren’t benefitting from our contacts either, so the matter may be moot.) If I’d not recommended we take our data out of VAN, this wouldn’t have been an issue.
The second issue was of uniformity: While organizers were generally reliable at entering contact data, they were inconsistent about how that data was entered. Most staff logged voter assessments on a five-point scale from 1–5, with “1” as a strong supporter, “2” as weak supporter, “3” as neutral, “4” and “5” as weak anti- and strong anti-, respectively. But organizers in some regions used a 7-point scale. And even among those using a five-point scale, there were several survey questions in use. Some variants designated the opposing candidate voters supported (e.g. “5K” = Strong Kerry; “5E” = Strong Edwards; “4N” = Weak Kucinich). Other variants were simply regionalized or medium-driven duplicates (e.g. “Will you support Howard Dean in the caucus? — Davenport/Bettendorf,” or ”Phone Bank Support Question”).
While there are plenty of good reasons to segregate different classes and methods of contact (especially for statistical analysis), it’s much easier for a data manager to accurately and quickly tally supporters when they just have to look at counts for one question and don’t need to do any additional math. But the more challenging problem is with conflict resolution: when a voter is tagged as a “1” by one survey question and a “5” by another, the data manager must look at the date of entry for each assessment, counting only the most recent in the supporter tally. In our case, this was a database management/permissions issue. There wasn’t a survey-question-mastering feature allowing for different question results to be caged together and counted in aggregate, so we should’ve set a policy identifying the one survey question and scale to be used by everyone and held folks to it.
The third issue was related to the second but a far deadlier error. To understand how many voters had been tagged with multiple types of conflicting support assessments, I exported the assessment data and re-figured it to show the most recent response (supporting/neutral/opposing), regardless which survey question had been selected. My results were shocking and at first inexplicable: I showed we had only about 30,000 likely caucus supporters.
The system was designed to include voters in counts based on all of their prior assessment results, even within a single survey question.
This meant a voter assessed in June as a weak Dean supporter, in August as a strong Dean supporter, in October as undecided, and a strong Kerry supporter by December, would have been included in current counts for each result. Despite being solidly in the Kerry column, the 130,000 supporters announced on the staff call my first night in the office included this voter, twice.
In data architecture lingo, the survey-answers object had a many-to-many relationship with the voter object. This design was likely selected in the interest of flexibility: users were expected to delete old assessments before they entered new ones. But users and administrators alike assumed the system would behave differently — to logically *replace* an old assessment when a new one comes in (as was/is VAN’s design). Though I wasn’t an architect of Dean’s voter database, I did recommend the vendor. Had I gotten over my paranoid instincts and agreed to use the VAN out the gate, we would’ve not experienced this issue.
I quadruple-checked my math, and re-ran my code to confirm: we were not going to win Iowa after all. Our own data, when properly interrogated, showed us behind. And we had very boldly celebrated numbers with our team I now knew were way, way off.
I quietly shared my findings with the two senior state leaders who had invited me in. They were shocked. They asked what I thought we could do. We were less than a month away from the caucuses, and we had thousands of out-of-state volunteers coming in for the Perfect Storm canvass blitz. Fixing the data and offering a mea culpa on the next all-staff call didn’t seem like a viable option.
If we’d caught the error earlier we would have seen supporters slipping away and adjusted our approach. But the only thing left to do at this point was to keep fighting, albeit with a sad, new understanding of our campaign’s impending doom. We may not have won, but this whole mess could have been avoided had we practiced data discipline, tested assumptions and run audits early and often.
Together, we broke the news to the National Campaign Manager outside a Des Moines brewpub in the frigid darkness, a light dusting of snow blowing around us. He instructed us to not tell anyone else what we had learned. I packed my stuff and headed back home to spend Christmas with my family.
The Iowa team carried on, believing the internal and external data showing us in the lead. But soon the polling would catch up with our own data and we would ultimately take third place in the caucus tally. Then there was the scream. A series of primary losses drained the last bit of momentum from the campaign and the Governor withdrew from the race.
While I doubt he ever learned about his campaign’s fatal data flaw, years later as DNC chairman Dr. Dean would inoculate every other Democratic campaign against suffering a similar fate by rolling out a 50-state VAN. This landmark vendor deal was just one facet of the Governor’s “50-State Strategy” that dramatically raised investment in local Democratic infrastructure across America and set the stage for Barack Obama’s historic 2008 primary and general election wins.
In the wake of Secretary Clinton’s recent loss, some question the role, value, and importance of data in modern campaigns. We all just learned an incredibly strong political data and analytics program couldn’t guarantee victory, but there are plenty of historical examples of weak data and technology programs dooming campaigns to failure.
Author’s Note:
Because I think Democratic institutional memory should focus on the campaigns’ relatable lessons, I’m writing under a pseudonym and haven’t used staff names. But it’s neither anonymous nor is there anything salacious or secret disclosed here: after the 2004 election I shared this experience with anyone who would listen. I think talking about our mistakes helped our party plot a new, better course for the next decade.
